A: Tor (like all current practical low-latency anonymity designs) fails when the attacker can see both ends of the communications channel. For example, suppose the attacker controls or watches the Tor relay you choose to enter the network, and also controls or watches the website you visit. In this case, the research community knows no practical low-latency design that can reliably stop the attacker from correlating volume and timing information on the two sides.

So, what should we do? Suppose the attacker controls, or can observe, C relays. Suppose there are N relays total. If you select new entry and exit relays each time you use the network, the attacker will be able to correlate all traffic you send with probability (c/n)2. But profiling is, for most users, as bad as being traced all the time: they want to do something often without an attacker noticing, and the attacker noticing once is as bad as the attacker noticing more often. Thus, choosing many random entries and exits gives the user no chance of escaping profiling by this kind of attacker.

The solution is "entry guards": each Tor client selects a few relays at random to use as entry points, and uses only those relays for her first hop. If those relays are not controlled or observed, the attacker can't win, ever, and the user is secure. If those relays are observed or controlled by the attacker, the attacker sees a larger fraction of the user's traffic — but still the user is no more profiled than before. Thus, the user has some chance (on the order of (n-c)/n) of avoiding profiling, whereas she had none before.